Abstract: Sensitivity analysis of hydrological models is a key step for model uncertainty quantification. It can identify the dominant parameters, reduce the model calibration uncertainty, and enhance the model optimization efficiency. However, how to effectively validate a model and identify the dominant parameters for a large-scale complex distributed hydrological model is a bottle-neck to achieve the parameters optimization. There are some shortcomings for classical approaches, e.g. time-consuming and high computation cost, to quantitatively assess the sensitivity of the multi-parameters complex hydrological model. For this reason, a new approach was applied in this paper, in which the support vector machine was used to construct the response surface (a surrogate model) at first. Then it integrated the SVM-based response surface with the Sobol method, i.e. the RSMSobol method, to achieve the quantification assessment of sensitivity for complex models. Taking the distributed time-variant gain model in the Huaihe River Basin as a case study, we selected three objective functions (i.e. water balance coefficient WB, Nash-Sutcliffe efficiency coefficient NS, and correlation coefficient RC) to assess the model as the output responses for sensitivity analysis. The results show that the RSMSobol method can not only achieve the quantification of the sensitivity, and also reduce the computational cost, with good accuracy compared to the classical approaches.

Key words: meta-modeling approach, response surface methodology, sensitivity analysis, support vector machines, Huaihe River Basin